Skip to the stories
Microsoft 365 & SharePoint

Why Localised Heat Is Pushing Data Centres Toward Rack-Level Cooling

  • Mark Bowen
  • 882 views

AI racks at 80–150 kW are creating hotspots room-level airflow can't fix. Here's why rear-door heat exchangers and hybrid cooling are gaining ground.

Why Localised Heat Is Pushing Data Centres Toward Rack-Level Cooling

The problem with AI infrastructure isn't that data halls are getting hotter overall. It's that heat is concentrating in specific places, and the cooling designs most facilities run were never built to handle that. Klaus Dafinger, Cooling Marketing Manager at Legrand, argues that operators now need to think about cooling at the rack, not just the room. For anyone planning capacity against GPU growth, that shift matters more than any single product choice.

Klaus Dafinger, Cooling Marketing Manager at Legrand

Why the room-level model is running out of road

Most legacy cooling strategies assumed fairly uniform rack densities. If you could push the right volume of air through the room and keep pressure balanced, you got predictable thermal performance. That assumption is breaking down. Many operators now run mixed estates, with older enterprise applications sitting next to GPU clusters drawing somewhere between 80 and 150 kW per rack. Those clusters throw off intense, localised hotspots that a room-wide airflow strategy handles poorly.

The conversation has accordingly moved on. The question is less about how to move more air through the data hall and more about removing heat at the point where it's generated.

Fan walls: useful, but not built for concentrated loads

Fan wall architectures have gained traction because they do a lot right. Multiple smaller fans in a dedicated wall give more uniform airflow, tighter pressure control, built-in redundancy and better energy efficiency than many traditional approaches. In a facility where thermal load is spread evenly across the white space, they work well.

The trouble is that AI rarely produces an even heat distribution. When GPU density is clustered in one area, a fan wall can't target airflow at individual high-density racks. It can only move air across the room. Dafinger's point, as I read it, is that fan walls remain a strong choice for many facilities, but their effectiveness drops when demand is concentrated in a few racks while the rest of the room runs at lower density.

Planning a fan wall installation also brings its own constraints. Teams should expect:

  • A large infrastructure footprint that needs dedicated wall space.
  • Performance that is best when thermal loads are relatively uniform.
  • Limited ability to direct airflow at individual high-density racks.
  • Success that depends on detailed airflow modelling and room design.
  • Typically, chilled water infrastructure and associated control systems.

From a room-temperature problem to a heat-source problem

The underlying shift is conceptual as much as mechanical. Cooling design is moving from managing room temperature to managing the heat output of specific IT loads. Operators are increasingly placing clusters of high-density AI servers alongside conventional enterprise kit, and those mixed-density environments put very different demands on the same cooling plant.

That's why many operators are looking for architectures that scale selectively. The goal is to target the racks and rows carrying the highest thermal risk without forcing change across the entire facility. Here's the thing: that approach lets you spend on cooling where the heat actually is, instead of over-provisioning the whole hall to cover a few hot racks.

Where rack-level cooling fits

Rear Door Heat Exchangers (RDHx) are the clearest example of this approach. A liquid-cooled heat exchanger is integrated into the cabinet's rear door. As warm exhaust air leaves the servers, its thermal energy transfers into the liquid circuit, and the conditioned air returns to the data hall. The heat is removed close to its source, rather than relying solely on room-level airflow.

For operators who need to support AI growth without committing to a full cooling redesign, this is a practical middle path. It can bridge conventional air cooling and more intensive liquid cooling. According to Legrand's framing, RDHx enables operators to:

  • Increase rack density where required, without upgrading cooling capacity across the entire data hall.
  • Align cooling capacity with actual heat generation, targeting only the highest-density racks.
  • Deploy cooling incrementally, so investment can follow business growth, AI adoption and server refresh cycles.
  • Minimise infrastructure disruption, since RDHx can often fit within existing rack footprints without major structural work.
  • Support mixed-density environments, cooling AI racks independently while conventional workloads stay on existing air-cooled infrastructure.
  • Improve operational resilience by localising cooling, which can reduce the impact of maintenance or faults compared with more centralised designs.

The incremental point deserves attention. Teams that can't justify a hall-wide retrofit can often start with the racks that are running hottest and expand from there.

Hybrid is becoming the default

Perhaps the biggest change is that cooling is no longer a choice between one technology and another. Many operators are combining room-level airflow management, rack-level cooling and direct liquid cooling, matched to workload requirements.

In practice, that often looks like this. Traditional enterprise applications keep running on established air cooling. High-density AI deployments may justify Direct-to-Chip (D2C) liquid cooling. In between sits a growing band of mixed-density environments, where RDHx can add capacity without wholesale redesign. Hybrid architectures let operators modernise in stages, weighing thermal performance, operational risk and capital spend while preparing for further AI growth.

What to watch next

I don't think a single cooling technology will define the next phase. The operators who do well will be the ones who match cooling capacity to workload demand precisely, and who can address concentrated heat without overhauling the whole hall. For many existing facilities, that means auditing where heat actually sits, separating AI racks from conventional ones, and pricing a phased rollout before committing to a full redesign.

For more on how RDHx technologies can support high-density AI environments and phased liquid cooling adoption, see the Legrand website.

Source: Mark Bowen · www.intelligentdatacentres.com

Neighbouring cells